{
  "schema_version": 1,
  "project": "sft",
  "title": "SFT 学到了行为，能力收益取决于起点",
  "recorded_at": "2026-10-04T12:09:45+08:00",
  "current": {
    "title": "SFT 学到了行为，能力收益取决于起点",
    "goal": "区分行为迁移、token 足迹与解题能力收益",
    "direction": "What Agentic SFT Buys: Rapid Behavioral Convergence, Base-Conditional Capability Gains",
    "status": "完整稿 · 结论有边界",
    "reliable": "行为迁移可测，dense 输入中位数下降 45—53%；小幅解决率差异仍未分辨。[[S1]]",
    "summary": "这篇研究问：用强代理的轨迹训练开放模型，究竟迁移了哪些东西？当前证据支持命令习惯、工具使用和错误恢复向教师接近，并伴随输入 token 足迹下降；解决率的收益则依赖学生起点与任务池。已有完整分析稿，但训练臂可比性、测量口径和最终投稿核对仍需收口。下一步先把最新锚点与正文主张逐项对齐。[[S1]][[S2]]",
    "next": "对齐最终正文、训练臂说明和统计口径，再关闭投稿检查缺口。[[S2]][[S5]]",
    "observed_at": "2026-09-06",
    "recorded_at": "2026-10-04T11:56:20+08:00",
    "recent_changes": [
      "本次页面以 2026-09-06 完整稿及对应记录替换旧 scaffold 印象。锚点显示为已补充；剂量图改为独立实验臂，避免误示同 run 曲线。",
      "来源 SHA 标识本次读取的文件内容；已有审计报告的结论只适用于其记录的版本和范围。未在本次网页制作中执行新的完整性/论文审查。"
    ],
    "constraints": [
      "原始结果、失败和unknown保留，不将缺判分当0。",
      "开始计算任务前核对当前进程、资源与准入收据。",
      "来源变动后核对受影响主张，网页不替代原始证据。",
      "不将s128与s256/s381画成严格同run剂量曲线。",
      "用户关闭的同条件重测与新代嫁接不自行重新启动。"
    ],
    "conflicts": [],
    "must_verify": [
      "当前作业、GPU与容器状态（本工作台未执行现场资源检查）。",
      "新增来源是否改变既有实验协议或科学准入。"
    ]
  },
  "documents": [
    {
      "id": "S1",
      "path": "paper/distill-fingerprint/tex/main.tex",
      "title": "现稿摘要与章节入口",
      "kind": "research",
      "observed_at": "2026-09-06 现稿",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "033790b2f28e199e122e496621afe54544eafdce2e1981d9920025ca381d2998",
      "priority": 2,
      "public": true,
      "sha256": "033790b2f28e199e122e496621afe54544eafdce2e1981d9920025ca381d2998",
      "reviewed_sha256": "033790b2f28e199e122e496621afe54544eafdce2e1981d9920025ca381d2998",
      "bytes": 5686,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/S1/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/S1/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "S2",
      "path": "paper/distill-fingerprint/PAPER_PLAN.md",
      "title": "贡献定位、锚点与关闭实验",
      "kind": "research",
      "observed_at": "2026-09-06 更新",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "d4b96d91203a84a8aacd3cdac2539750773e8aebc34492c58ca5d6dcd7b0697e",
      "priority": 2,
      "public": true,
      "sha256": "d4b96d91203a84a8aacd3cdac2539750773e8aebc34492c58ca5d6dcd7b0697e",
      "reviewed_sha256": "d4b96d91203a84a8aacd3cdac2539750773e8aebc34492c58ca5d6dcd7b0697e",
      "bytes": 24742,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/S2/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/S2/source.md",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "S3",
      "path": "paper/distill-fingerprint/tex/tables/tab1_arms.tex",
      "title": "主表：分数、输入中位数与 Bash-JS",
      "kind": "research",
      "observed_at": "2026-09-06 当前文件",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "49818f7b9d95804c8cbcad1363ee17f1b9f21f0ec202568c5ea7d8027e4e7ac7",
      "priority": 2,
      "public": true,
      "sha256": "49818f7b9d95804c8cbcad1363ee17f1b9f21f0ec202568c5ea7d8027e4e7ac7",
      "reviewed_sha256": "49818f7b9d95804c8cbcad1363ee17f1b9f21f0ec202568c5ea7d8027e4e7ac7",
      "bytes": 2663,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/S3/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/S3/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "S4",
      "path": "paper/distill-fingerprint/findings.md",
      "title": "run 差异、锚点与评分限制",
      "kind": "research",
      "observed_at": "截至 2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "84f0cc4512a7c02d4047df72e2b9ba1fc352695ed9e3f130bb6115f840f12034",
      "priority": 2,
      "public": true,
      "sha256": "84f0cc4512a7c02d4047df72e2b9ba1fc352695ed9e3f130bb6115f840f12034",
      "reviewed_sha256": "84f0cc4512a7c02d4047df72e2b9ba1fc352695ed9e3f130bb6115f840f12034",
      "bytes": 24548,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/S4/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/S4/source.md",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "S5",
      "path": "paper/distill-fingerprint/tex/sections/09_limitations.tex",
      "title": "现稿 Limitations",
      "kind": "research",
      "observed_at": "当前稿件",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "9b04b60a149f76f17836cda5db7d0ede47a7f2bf1adfe82ad86fc24c1e313e57",
      "priority": 2,
      "public": true,
      "sha256": "9b04b60a149f76f17836cda5db7d0ede47a7f2bf1adfe82ad86fc24c1e313e57",
      "reviewed_sha256": "9b04b60a149f76f17836cda5db7d0ede47a7f2bf1adfe82ad86fc24c1e313e57",
      "bytes": 2753,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/S5/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/S5/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "S6",
      "path": "paper/distill-fingerprint/CLAIMS_FROM_RESULTS.md",
      "title": "早期 result-to-claim（历史版本）",
      "kind": "research",
      "observed_at": "2026-09-02",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "415be75a7d8c48ee154b3ad6afc41715c608a2b517b85993c8b2614f84df36c4",
      "priority": 2,
      "public": true,
      "sha256": "415be75a7d8c48ee154b3ad6afc41715c608a2b517b85993c8b2614f84df36c4",
      "reviewed_sha256": "415be75a7d8c48ee154b3ad6afc41715c608a2b517b85993c8b2614f84df36c4",
      "bytes": 9099,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/S6/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/S6/source.md",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "S7",
      "path": "paper/distill-fingerprint/CITATION_AUDIT.md",
      "title": "引用审计与已应用修订",
      "kind": "research",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "c22361b34fe7aa618d098c42b813c85211be459548b704fe4996e8ae7f3edec1",
      "priority": 2,
      "public": true,
      "sha256": "c22361b34fe7aa618d098c42b813c85211be459548b704fe4996e8ae7f3edec1",
      "reviewed_sha256": "c22361b34fe7aa618d098c42b813c85211be459548b704fe4996e8ae7f3edec1",
      "bytes": 5219,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/S7/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/S7/source.md",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "EXPERIMENT_LOG",
      "path": "paper/distill-fingerprint/EXPERIMENT_LOG.md",
      "title": "实验与推断记录",
      "kind": "history",
      "observed_at": null,
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "44a2e9c93736a5035551a07d26c632b06cd48cf0af89af691c28f1d390844a61",
      "priority": 2,
      "public": true,
      "sha256": "44a2e9c93736a5035551a07d26c632b06cd48cf0af89af691c28f1d390844a61",
      "reviewed_sha256": "44a2e9c93736a5035551a07d26c632b06cd48cf0af89af691c28f1d390844a61",
      "bytes": 6676,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/EXPERIMENT_LOG/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/EXPERIMENT_LOG/source.md",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "NARRATIVE",
      "path": "paper/distill-fingerprint/NARRATIVE_REPORT.md",
      "title": "完整研究叙述",
      "kind": "research",
      "observed_at": null,
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "f3cfeee75fb4b3ffbb626f853762ebf3a132a5e163575ba4647a6e094114caa3",
      "priority": 2,
      "public": true,
      "sha256": "f3cfeee75fb4b3ffbb626f853762ebf3a132a5e163575ba4647a6e094114caa3",
      "reviewed_sha256": "f3cfeee75fb4b3ffbb626f853762ebf3a132a5e163575ba4647a6e094114caa3",
      "bytes": 38408,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/NARRATIVE/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/NARRATIVE/source.md",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "NOVELTY",
      "path": "paper/distill-fingerprint/NOVELTY_REPORT.md",
      "title": "创新定位报告",
      "kind": "research",
      "observed_at": null,
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "ca356ccc689a970352a412ae38230bef6d8e26d207d674aedb259987bbfa3956",
      "priority": 2,
      "public": true,
      "sha256": "ca356ccc689a970352a412ae38230bef6d8e26d207d674aedb259987bbfa3956",
      "reviewed_sha256": "ca356ccc689a970352a412ae38230bef6d8e26d207d674aedb259987bbfa3956",
      "bytes": 12040,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/NOVELTY/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/NOVELTY/source.md",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "STATS",
      "path": "paper/distill-fingerprint/analysis/stats/stats_fixes.md",
      "title": "统计口径与修订",
      "kind": "definition",
      "observed_at": null,
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "468d07b12a066b13333b6e270e9d14bbfab881be988b2255deb73cf95e5822ea",
      "priority": 1,
      "public": true,
      "sha256": "468d07b12a066b13333b6e270e9d14bbfab881be988b2255deb73cf95e5822ea",
      "reviewed_sha256": "468d07b12a066b13333b6e270e9d14bbfab881be988b2255deb73cf95e5822ea",
      "bytes": 50907,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/STATS/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/STATS/source.md",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "CLAIM_CODE",
      "path": "paper/distill-fingerprint/build_claims.py",
      "title": "主张证据提取器",
      "kind": "code",
      "observed_at": null,
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "65f4b3a0707414c214c301d5f6fb446ebbe4beb54f89c4aeffd43d7f8189e598",
      "priority": 2,
      "public": true,
      "sha256": "65f4b3a0707414c214c301d5f6fb446ebbe4beb54f89c4aeffd43d7f8189e598",
      "reviewed_sha256": "65f4b3a0707414c214c301d5f6fb446ebbe4beb54f89c4aeffd43d7f8189e598",
      "bytes": 9020,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/CLAIM_CODE/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/CLAIM_CODE/source.py",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "TEX_01_INTRO",
      "path": "paper/distill-fingerprint/tex/sections/01_intro.tex",
      "title": "01_intro · 完整稿件源码",
      "kind": "manuscript",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "af4e1899cc53fa4dfc7562dd77c03d94c0f2111abde63a142e1bfc5a62f25d47",
      "priority": 2,
      "public": true,
      "sha256": "af4e1899cc53fa4dfc7562dd77c03d94c0f2111abde63a142e1bfc5a62f25d47",
      "reviewed_sha256": "af4e1899cc53fa4dfc7562dd77c03d94c0f2111abde63a142e1bfc5a62f25d47",
      "bytes": 5199,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_01_INTRO/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_01_INTRO/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "TEX_02_RELATED",
      "path": "paper/distill-fingerprint/tex/sections/02_related.tex",
      "title": "02_related · 完整稿件源码",
      "kind": "manuscript",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "733fff0236a96a4b5d52c8f1427a4d9a7608ae146ea39728ffded5c84ec536c4",
      "priority": 2,
      "public": true,
      "sha256": "733fff0236a96a4b5d52c8f1427a4d9a7608ae146ea39728ffded5c84ec536c4",
      "reviewed_sha256": "733fff0236a96a4b5d52c8f1427a4d9a7608ae146ea39728ffded5c84ec536c4",
      "bytes": 3776,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_02_RELATED/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_02_RELATED/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "TEX_03_CORPUS",
      "path": "paper/distill-fingerprint/tex/sections/03_corpus.tex",
      "title": "03_corpus · 完整稿件源码",
      "kind": "manuscript",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "733c9d6ed65ad7fe70ce41b0439ca7870a1b14cade7447c0e1168aa19f8f7bfe",
      "priority": 2,
      "public": true,
      "sha256": "733c9d6ed65ad7fe70ce41b0439ca7870a1b14cade7447c0e1168aa19f8f7bfe",
      "reviewed_sha256": "733c9d6ed65ad7fe70ce41b0439ca7870a1b14cade7447c0e1168aa19f8f7bfe",
      "bytes": 4909,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_03_CORPUS/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_03_CORPUS/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "TEX_04_SETUP",
      "path": "paper/distill-fingerprint/tex/sections/04_setup.tex",
      "title": "04_setup · 完整稿件源码",
      "kind": "manuscript",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "d65ea24a19040f5cf0bf51ae27fb071ad104cc450ddcc846c2f144b501b16681",
      "priority": 2,
      "public": true,
      "sha256": "d65ea24a19040f5cf0bf51ae27fb071ad104cc450ddcc846c2f144b501b16681",
      "reviewed_sha256": "d65ea24a19040f5cf0bf51ae27fb071ad104cc450ddcc846c2f144b501b16681",
      "bytes": 3662,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_04_SETUP/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_04_SETUP/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "TEX_05_INHERIT",
      "path": "paper/distill-fingerprint/tex/sections/05_inherit.tex",
      "title": "05_inherit · 完整稿件源码",
      "kind": "manuscript",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "8df544cce4b4c7d5ef1f5e4994636782287c4b6c6971b09d63beccfbe15ebf84",
      "priority": 2,
      "public": true,
      "sha256": "8df544cce4b4c7d5ef1f5e4994636782287c4b6c6971b09d63beccfbe15ebf84",
      "reviewed_sha256": "8df544cce4b4c7d5ef1f5e4994636782287c4b6c6971b09d63beccfbe15ebf84",
      "bytes": 8313,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_05_INHERIT/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_05_INHERIT/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "TEX_06_COSTS",
      "path": "paper/distill-fingerprint/tex/sections/06_costs.tex",
      "title": "06_costs · 完整稿件源码",
      "kind": "manuscript",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "036fcff14faa13466439dbcb2b2496721d335ae0efa92c3c5316b68137a27799",
      "priority": 2,
      "public": true,
      "sha256": "036fcff14faa13466439dbcb2b2496721d335ae0efa92c3c5316b68137a27799",
      "reviewed_sha256": "036fcff14faa13466439dbcb2b2496721d335ae0efa92c3c5316b68137a27799",
      "bytes": 2692,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_06_COSTS/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_06_COSTS/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "TEX_07_MEASUREMENT",
      "path": "paper/distill-fingerprint/tex/sections/07_measurement.tex",
      "title": "07_measurement · 完整稿件源码",
      "kind": "manuscript",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "aa8cae2fee4d183d855dffd6b0f155137ad71e1cfc4a506cfeb82fcd199321b4",
      "priority": 2,
      "public": true,
      "sha256": "aa8cae2fee4d183d855dffd6b0f155137ad71e1cfc4a506cfeb82fcd199321b4",
      "reviewed_sha256": "aa8cae2fee4d183d855dffd6b0f155137ad71e1cfc4a506cfeb82fcd199321b4",
      "bytes": 2646,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_07_MEASUREMENT/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_07_MEASUREMENT/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "TEX_08_CONCLUSION",
      "path": "paper/distill-fingerprint/tex/sections/08_conclusion.tex",
      "title": "08_conclusion · 完整稿件源码",
      "kind": "manuscript",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "77040cf4940a09d65409e8333779d12567e0e710de3787a9ffdef9732c81c6ec",
      "priority": 2,
      "public": true,
      "sha256": "77040cf4940a09d65409e8333779d12567e0e710de3787a9ffdef9732c81c6ec",
      "reviewed_sha256": "77040cf4940a09d65409e8333779d12567e0e710de3787a9ffdef9732c81c6ec",
      "bytes": 945,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_08_CONCLUSION/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_08_CONCLUSION/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "TEX_10_ETHICS",
      "path": "paper/distill-fingerprint/tex/sections/10_ethics.tex",
      "title": "10_ethics · 完整稿件源码",
      "kind": "manuscript",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "e131234c0d9b40191b52b08a3bf0ede3d040f22fd9d29a9d50346a48e8d86c52",
      "priority": 2,
      "public": true,
      "sha256": "e131234c0d9b40191b52b08a3bf0ede3d040f22fd9d29a9d50346a48e8d86c52",
      "reviewed_sha256": "e131234c0d9b40191b52b08a3bf0ede3d040f22fd9d29a9d50346a48e8d86c52",
      "bytes": 1059,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_10_ETHICS/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_10_ETHICS/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "TEX_A_EVALUATION_NOISE",
      "path": "paper/distill-fingerprint/tex/appendix/A_evaluation_noise.tex",
      "title": "A_evaluation_noise · 完整稿件源码",
      "kind": "manuscript",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "07f2474848cbd65065894380eb787127d8768fa5221316e2aaf75d9cbae96007",
      "priority": 2,
      "public": true,
      "sha256": "07f2474848cbd65065894380eb787127d8768fa5221316e2aaf75d9cbae96007",
      "reviewed_sha256": "07f2474848cbd65065894380eb787127d8768fa5221316e2aaf75d9cbae96007",
      "bytes": 7682,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_A_EVALUATION_NOISE/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_A_EVALUATION_NOISE/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "TEX_B_FINGERPRINT_ROBUSTNESS",
      "path": "paper/distill-fingerprint/tex/appendix/B_fingerprint_robustness.tex",
      "title": "B_fingerprint_robustness · 完整稿件源码",
      "kind": "manuscript",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "7fc22a35894c1ce6c30c8634309260d4763e7d1f047c9eb696004cef316967a3",
      "priority": 2,
      "public": true,
      "sha256": "7fc22a35894c1ce6c30c8634309260d4763e7d1f047c9eb696004cef316967a3",
      "reviewed_sha256": "7fc22a35894c1ce6c30c8634309260d4763e7d1f047c9eb696004cef316967a3",
      "bytes": 11086,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_B_FINGERPRINT_ROBUSTNESS/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_B_FINGERPRINT_ROBUSTNESS/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "TEX_C_EXTRAS",
      "path": "paper/distill-fingerprint/tex/appendix/C_extras.tex",
      "title": "C_extras · 完整稿件源码",
      "kind": "manuscript",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "b315ecb1ec4af7617bff810815e79a28c0b527de2371e5c4252e282a4fee865a",
      "priority": 2,
      "public": true,
      "sha256": "b315ecb1ec4af7617bff810815e79a28c0b527de2371e5c4252e282a4fee865a",
      "reviewed_sha256": "b315ecb1ec4af7617bff810815e79a28c0b527de2371e5c4252e282a4fee865a",
      "bytes": 8076,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_C_EXTRAS/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_C_EXTRAS/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "TEX_D_REPRODUCIBILITY",
      "path": "paper/distill-fingerprint/tex/appendix/D_reproducibility.tex",
      "title": "D_reproducibility · 完整稿件源码",
      "kind": "manuscript",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "de2b08219d1092ca8203150b42c266d8019e8d3c635e1c66cce232cb0678c22c",
      "priority": 2,
      "public": true,
      "sha256": "de2b08219d1092ca8203150b42c266d8019e8d3c635e1c66cce232cb0678c22c",
      "reviewed_sha256": "de2b08219d1092ca8203150b42c266d8019e8d3c635e1c66cce232cb0678c22c",
      "bytes": 2712,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_D_REPRODUCIBILITY/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_D_REPRODUCIBILITY/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "TEX_E_ANCHORS",
      "path": "paper/distill-fingerprint/tex/appendix/E_anchors.tex",
      "title": "E_anchors · 完整稿件源码",
      "kind": "manuscript",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "430c9a9edfc29e434c082e26e3f44756149a6ca7efa3e46241078b391939cca8",
      "priority": 2,
      "public": true,
      "sha256": "430c9a9edfc29e434c082e26e3f44756149a6ca7efa3e46241078b391939cca8",
      "reviewed_sha256": "430c9a9edfc29e434c082e26e3f44756149a6ca7efa3e46241078b391939cca8",
      "bytes": 2798,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_E_ANCHORS/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/TEX_E_ANCHORS/source.tex",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986"
    },
    {
      "id": "CURRENT_PDF",
      "path": "paper/distill-fingerprint/tex/main.pdf",
      "title": "当前PDF",
      "kind": "artifact",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "expected_sha256": "59444261e49a0f43f0b8d25b9786c8d2822e1faaf0c549b7bdbea963c8d6f4d0",
      "priority": 2,
      "public": true,
      "sha256": "59444261e49a0f43f0b8d25b9786c8d2822e1faaf0c549b7bdbea963c8d6f4d0",
      "reviewed_sha256": "59444261e49a0f43f0b8d25b9786c8d2822e1faaf0c549b7bdbea963c8d6f4d0",
      "bytes": 613234,
      "read_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/CURRENT_PDF/",
      "raw_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/CURRENT_PDF/source.pdf",
      "source_changed_since_review": false,
      "release_version": "v2-eb1a844c9b153986",
      "text_url": "https://sotalabs.cn/papers/_releases/v2-eb1a844c9b153986/sft/library/CURRENT_PDF/text.txt"
    }
  ],
  "issues": [
    {
      "id": "SFT-I001",
      "title": "训练臂和正文主张的可比性",
      "category": "证据",
      "priority": 1,
      "status": "待处理",
      "description": "影响“剂量收敛”和起点条件下的能力解释。",
      "evidence": "s128 与 s256/s381 的 run/context 不同。[[S4]]",
      "attempts": [
        "参见关联证据与历史路线；本次导入没有新增科研尝试。"
      ],
      "unknowns": "具体缺口与现场状态仍需核对，不能从旧记录推断当前进程。",
      "next_action": "逐项对齐正文、表注、图注与训练配置；去除严格受控曲线暗示。",
      "done": "每个训练臂的条件、比较对象和推断范围都能定位；无越界的剂量因果句。",
      "dependencies": [],
      "first_discovered_at": null,
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "updated_at": "2026-10-04T11:56:20+08:00",
      "observed_at": "2026-09-06",
      "sources": [
        "S4"
      ],
      "resolved_evidence": [],
      "closure_reason": null,
      "history": [],
      "journals": [
        "2026-09-02",
        "2026-09-03",
        "2026-09-05"
      ]
    },
    {
      "id": "SFT-I002",
      "title": "扰动、异常和标注的范围",
      "category": "测量",
      "priority": 2,
      "status": "待处理",
      "description": "影响可靠性及结束语行为结论。",
      "evidence": "温度扰动不等于同条件重测；结束语没有人工标注。[[S2]][[S5]]",
      "attempts": [
        "参见关联证据与历史路线；本次导入没有新增科研尝试。"
      ],
      "unknowns": "具体缺口与现场状态仍需核对，不能从旧记录推断当前进程。",
      "next_action": "检查全部量尺说明，明确主评分规则与补充异常读数。",
      "done": "温度、异常、标注来源在每个相应结果旁有说明，保持既有关闭实验决定。",
      "dependencies": [],
      "first_discovered_at": null,
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "updated_at": "2026-10-04T11:56:20+08:00",
      "observed_at": "2026-09-06",
      "sources": [
        "S2",
        "S5"
      ],
      "resolved_evidence": [],
      "closure_reason": null,
      "history": [],
      "journals": [
        "2026-09-06",
        "2026-09-06"
      ]
    },
    {
      "id": "SFT-I003",
      "title": "历史待办与最终交付状态",
      "category": "写作",
      "priority": 3,
      "status": "待处理",
      "description": "影响合作者判断哪些内容仍缺，以及投稿准备度。",
      "evidence": "锚点已落地，旧 scaffold README 与早期待办尚未完全同步。[[S1]][[S2]]",
      "attempts": [
        "参见关联证据与历史路线；本次导入没有新增科研尝试。"
      ],
      "unknowns": "具体缺口与现场状态仍需核对，不能从旧记录推断当前进程。",
      "next_action": "建立最终主张—来源清单，完成稿件与引用/匿名/页数检查。",
      "done": "待办逐项有关闭证据；最终审查报告与具体稿件版本绑定。",
      "dependencies": [],
      "first_discovered_at": null,
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "updated_at": "2026-10-04T11:56:20+08:00",
      "observed_at": "2026-09-06",
      "sources": [
        "S1",
        "S2"
      ],
      "resolved_evidence": [],
      "closure_reason": null,
      "history": [],
      "journals": [
        "2026-09-06",
        "2026-09-06"
      ]
    },
    {
      "id": "SFT-I103",
      "title": "同条件重测 / 新代模型嫁接 · 历史路线",
      "category": "历史",
      "priority": 9,
      "status": "关闭",
      "description": "记录中由用户停止重测、放弃嫁接。",
      "evidence": "记录中由用户停止重测、放弃嫁接。",
      "attempts": [
        "记录中由用户停止重测、放弃嫁接。"
      ],
      "unknowns": "关闭仅适用于该固定路线，不证明同类方法全部无效。",
      "next_action": "保留历史，不通过改名重新启动。",
      "done": "保留测量限制，不擅自把关闭路线加入近期执行队列。[[S2]]",
      "dependencies": [],
      "first_discovered_at": null,
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "updated_at": "2026-10-04T11:56:20+08:00",
      "observed_at": "2026-09-06",
      "sources": [
        "S2"
      ],
      "resolved_evidence": [],
      "closure_reason": "保留测量限制，不擅自把关闭路线加入近期执行队列。[[S2]]",
      "history": [],
      "journals": [
        "2026-09-06",
        "2026-09-06"
      ]
    },
    {
      "id": "SFT-I090",
      "title": "教师同任务锚点不足（历史缺口已补充）",
      "category": "证据",
      "priority": 3,
      "status": "已解决",
      "description": "原先需要教师与学生在同一任务上的行为比较。",
      "evidence": "09-06计划记录359 Pro锚点与345 held-out验证分析已落地。[[S2]][[S4]]",
      "attempts": [
        "原先result-to-claim提出补充held-out锚点。",
        "后续锚点分析接入正文和附录。"
      ],
      "unknowns": "解决该缺口不证明跨教师/跨harness泛化。",
      "next_action": "保留来源，最终稿仍核对相应比较范围。",
      "done": "对应锚点分析与正文/附录材料存在。",
      "dependencies": [],
      "first_discovered_at": "2026-09-02",
      "recorded_at": "2026-10-04T12:09:45+08:00",
      "updated_at": "2026-10-04T12:09:45+08:00",
      "observed_at": "2026-09-06",
      "sources": [
        "S2",
        "S4"
      ],
      "resolved_evidence": [
        "S2",
        "S4"
      ],
      "closure_reason": null,
      "history": []
    }
  ],
  "tasks": [
    {
      "id": "SFT-T001",
      "title": "先核对最终主张与来源",
      "status": "待处理",
      "action": "以现稿为入口，将摘要、结论、表图与最新锚点逐条映射。",
      "done": "产出一份可追溯清单；若某句缺证据，先收窄或删除。",
      "dependencies": [],
      "issues": [
        "SFT-I001"
      ],
      "plan_kind": "网页整理的下一步；执行前按原研究协议核对",
      "observed_at": "2026-09-06"
    },
    {
      "id": "SFT-T002",
      "title": "再消除比较与测量歧义",
      "status": "待处理",
      "action": "依赖第一步清单，集中修订训练 run、温度扰动、异常及人工标注缺口的表述。",
      "done": "产出一致的正文和图注；需要新实验的强主张改为限制，不自动扩实验。",
      "dependencies": [
        "SFT-T001"
      ],
      "issues": [
        "SFT-I002"
      ],
      "plan_kind": "网页整理的下一步；执行前按原研究协议核对",
      "observed_at": "2026-09-06"
    },
    {
      "id": "SFT-T003",
      "title": "最后核对交付版本",
      "status": "待处理",
      "action": "完成前两步后，对目标会议规则、论断、引用、匿名与 PDF 做最终检查。",
      "done": "检查与稿件哈希绑定，仍有缺口就保留 draft；全部关闭才判断可交付。",
      "dependencies": [
        "SFT-T002"
      ],
      "issues": [
        "SFT-I003"
      ],
      "plan_kind": "网页整理的下一步；执行前按原研究协议核对",
      "observed_at": "2026-09-06"
    }
  ],
  "experiments": [
    {
      "id": "SFT-E001",
      "title": "dense native 与 SFT",
      "purpose": "核对当前主张的证据范围；详见原文设置。",
      "setup": "Bash-JS 为 log₂ Jensen–Shannon 距离，比较 Pro 上命令头词分布与教师参考，越小越近。输入为每题累计输入 token 的中位数，单位 M；解决率使用 Pro 731、Verified 500 的固定分母，异常 trial 在主表算未解出。[[S3]]",
      "result": "Bash-JS：0.202 → 0.069 / 0.053 / 0.050；Pro 分数 50.1 → 49.7 / 52.8 / 51.0%。",
      "interpretation": "行为明显靠近教师，但分数没有同样单调的关系。顺序为 s128/s256/s381，不作同 run 因果曲线。[[S3]][[S4]]",
      "technical_status": "已记录",
      "scientific_status": "限于来源范围",
      "cost": "按来源的输入、输出、服务时间与GPU小时口径阅读，未知不补零。",
      "observed_at": "2026-09-06",
      "sources": [
        "S3",
        "S4"
      ],
      "issues": [
        "SFT-I001"
      ]
    },
    {
      "id": "SFT-E002",
      "title": "裸基座对照",
      "purpose": "核对当前主张的证据范围；详见原文设置。",
      "setup": "Bash-JS 为 log₂ Jensen–Shannon 距离，比较 Pro 上命令头词分布与教师参考，越小越近。输入为每题累计输入 token 的中位数，单位 M；解决率使用 Pro 731、Verified 500 的固定分母，异常 trial 在主表算未解出。[[S3]]",
      "result": "Pro 35.4→44.3%，Δ=+8.89 pp；Verified 52.0→55.4%，Δ=+3.40 pp。",
      "interpretation": "一个裸基座的 Pro 收益可见；Verified 的差值在原审查中仍不确定。[[S3]][[S6]]",
      "technical_status": "已记录",
      "scientific_status": "限于来源范围",
      "cost": "按来源的输入、输出、服务时间与GPU小时口径阅读，未知不补零。",
      "observed_at": "2026-09-06",
      "sources": [
        "S3",
        "S6"
      ],
      "issues": []
    },
    {
      "id": "SFT-E003",
      "title": "已后训练 dense 臂",
      "purpose": "核对当前主张的证据范围；详见原文设置。",
      "setup": "Bash-JS 为 log₂ Jensen–Shannon 距离，比较 Pro 上命令头词分布与教师参考，越小越近。输入为每题累计输入 token 的中位数，单位 M；解决率使用 Pro 731、Verified 500 的固定分母，异常 trial 在主表算未解出。[[S3]]",
      "result": "输入中位数从 3.27 M 降到 1.69 / 1.52 / 1.75 M token/题。",
      "interpretation": "输入足迹变化不能简化为能力变化，也不能直接当成净费用。[[S3]]",
      "technical_status": "已记录",
      "scientific_status": "限于来源范围",
      "cost": "按来源的输入、输出、服务时间与GPU小时口径阅读，未知不补零。",
      "observed_at": "2026-09-06",
      "sources": [
        "S3"
      ],
      "issues": []
    },
    {
      "id": "SFT-E004",
      "title": "同任务锚点",
      "purpose": "核对当前主张的证据范围；详见原文设置。",
      "setup": "Bash-JS 为 log₂ Jensen–Shannon 距离，比较 Pro 上命令头词分布与教师参考，越小越近。输入为每题累计输入 token 的中位数，单位 M；解决率使用 Pro 731、Verified 500 的固定分母，异常 trial 在主表算未解出。[[S3]]",
      "result": "359 Pro 教师任务、345 held-out 验证任务进入分析和附录。",
      "interpretation": "原先缺少锚点的问题已有补充，旧待办不能照搬；跨任务行为转移仍受仓库重叠约束。[[S4]][[S5]]",
      "technical_status": "已记录",
      "scientific_status": "限于来源范围",
      "cost": "按来源的输入、输出、服务时间与GPU小时口径阅读，未知不补零。",
      "observed_at": "2026-09-06",
      "sources": [
        "S4",
        "S5"
      ],
      "issues": [
        "SFT-I001",
        "SFT-I002"
      ]
    },
    {
      "id": "SFT-E005",
      "title": "失败结束语标注",
      "purpose": "核对当前主张的证据范围；详见原文设置。",
      "setup": "Bash-JS 为 log₂ Jensen–Shannon 距离，比较 Pro 上命令头词分布与教师参考，越小越近。输入为每题累计输入 token 的中位数，单位 M；解决率使用 Pro 731、Verified 500 的固定分母，异常 trial 在主表算未解出。[[S3]]",
      "result": "110 个 regex-negative 失败结束语未标出承认失败；标注来自模型。",
      "interpretation": "只能解释完成措辞；没有人工标注，不能把 regex 指标称为诚实性评价。[[S5]]",
      "technical_status": "已记录",
      "scientific_status": "限于来源范围",
      "cost": "按来源的输入、输出、服务时间与GPU小时口径阅读，未知不补零。",
      "observed_at": "2026-09-06",
      "sources": [
        "S5"
      ],
      "issues": [
        "SFT-I002"
      ]
    }
  ],
  "decisions": [
    {
      "id": "SFT-D001",
      "title": "语料变体 v13",
      "question": "这条路线是否仍适合承担当前研究主张？",
      "alternatives": [
        "继续作为主方法",
        "保留为诊断/基线",
        "关闭或等待准入"
      ],
      "observation": "同时改变筛选与训练因素，分数与行为响应不完全一致。",
      "choice": "保留为复合变体，不当作某一筛选规则的单独效果。[[S5]]",
      "reason": "以原记录中的证据与门槛为依据，不另造因果解释。",
      "expected": "仅按明确贡献范围继续；未执行计划不能当结果。",
      "actual": "同时改变筛选与训练因素，分数与行为响应不完全一致。",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "sources": [
        "S5"
      ],
      "issues": [
        "SFT-I002"
      ]
    },
    {
      "id": "SFT-D002",
      "title": "教师同任务锚点",
      "question": "这条路线是否仍适合承担当前研究主张？",
      "alternatives": [
        "继续作为主方法",
        "保留为诊断/基线",
        "关闭或等待准入"
      ],
      "observation": "后续补齐 Pro 与未训练验证任务分析。",
      "choice": "把早期稿缺口更新为已补充，而不是继续展示 DATA_NEEDED。[[S2]][[S4]]",
      "reason": "以原记录中的证据与门槛为依据，不另造因果解释。",
      "expected": "仅按明确贡献范围继续；未执行计划不能当结果。",
      "actual": "后续补齐 Pro 与未训练验证任务分析。",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "sources": [
        "S2",
        "S4"
      ],
      "issues": [
        "SFT-I001",
        "SFT-I002",
        "SFT-I003",
        "SFT-I103"
      ]
    },
    {
      "id": "SFT-D003",
      "title": "同条件重测 / 新代模型嫁接",
      "question": "这条路线是否仍适合承担当前研究主张？",
      "alternatives": [
        "继续作为主方法",
        "保留为诊断/基线",
        "关闭或等待准入"
      ],
      "observation": "记录中由用户停止重测、放弃嫁接。",
      "choice": "保留测量限制，不擅自把关闭路线加入近期执行队列。[[S2]]",
      "reason": "以原记录中的证据与门槛为依据，不另造因果解释。",
      "expected": "仅按明确贡献范围继续；未执行计划不能当结果。",
      "actual": "记录中由用户停止重测、放弃嫁接。",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "sources": [
        "S2"
      ],
      "issues": [
        "SFT-I002",
        "SFT-I003",
        "SFT-I103"
      ]
    }
  ],
  "events": [
    {
      "id": "SFT-H001",
      "date": "2026-09-02",
      "title": "建立贡献与证据边界",
      "kind": "补录",
      "occurred_at": null,
      "observed_at": "2026-09-02",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "fields": {
        "当天结论": [
          "建立贡献与证据边界"
        ],
        "完成的工作": [
          "早期 result-to-claim 判 PARTIAL；完整初稿进入写作。[[S6]][[S4]]"
        ],
        "新证据与结果": [
          "早期 result-to-claim 判 PARTIAL；完整初稿进入写作。[[S6]][[S4]]"
        ],
        "判断与决策变化": [
          "补录仅归纳原记录；详细决定与历史结果见关联原文。"
        ],
        "问题变化": [
          "关联问题用于追踪背景；这里不推断当天已解决。"
        ],
        "论文变化": [
          "具体稿件版本见论文栏目；本条历史补录不推断当天改稿。"
        ],
        "下一步与交接": [
          "这是历史记录，当前行动以最新概览和接手包为准。"
        ]
      },
      "sources": [
        "S6",
        "S4"
      ],
      "issues": [
        "SFT-I001"
      ],
      "experiments": [
        "SFT-E001",
        "SFT-E002",
        "SFT-E004"
      ],
      "decisions": [
        "SFT-D002"
      ],
      "tasks": [],
      "artifacts": [],
      "corrects": null
    },
    {
      "id": "SFT-H002",
      "date": "2026-09-03",
      "title": "发现 run 可比性问题",
      "kind": "补录",
      "occurred_at": null,
      "observed_at": "2026-09-03",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "fields": {
        "当天结论": [
          "发现 run 可比性问题"
        ],
        "完成的工作": [
          "s128 不能与另两个点当成同 run 训练曲线。[[S4]]"
        ],
        "新证据与结果": [
          "s128 不能与另两个点当成同 run 训练曲线。[[S4]]"
        ],
        "判断与决策变化": [
          "补录仅归纳原记录；详细决定与历史结果见关联原文。"
        ],
        "问题变化": [
          "关联问题用于追踪背景；这里不推断当天已解决。"
        ],
        "论文变化": [
          "具体稿件版本见论文栏目；本条历史补录不推断当天改稿。"
        ],
        "下一步与交接": [
          "这是历史记录，当前行动以最新概览和接手包为准。"
        ]
      },
      "sources": [
        "S4"
      ],
      "issues": [
        "SFT-I001"
      ],
      "experiments": [
        "SFT-E001",
        "SFT-E004"
      ],
      "decisions": [
        "SFT-D002"
      ],
      "tasks": [],
      "artifacts": [],
      "corrects": null
    },
    {
      "id": "SFT-H003",
      "date": "2026-09-05",
      "title": "验证锚点完成",
      "kind": "补录",
      "occurred_at": null,
      "observed_at": "2026-09-05",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "fields": {
        "当天结论": [
          "验证锚点完成"
        ],
        "完成的工作": [
          "八臂 345 任务的严格与宽松评分需分别读。[[S4]]"
        ],
        "新证据与结果": [
          "八臂 345 任务的严格与宽松评分需分别读。[[S4]]"
        ],
        "判断与决策变化": [
          "补录仅归纳原记录；详细决定与历史结果见关联原文。"
        ],
        "问题变化": [
          "关联问题用于追踪背景；这里不推断当天已解决。"
        ],
        "论文变化": [
          "具体稿件版本见论文栏目；本条历史补录不推断当天改稿。"
        ],
        "下一步与交接": [
          "这是历史记录，当前行动以最新概览和接手包为准。"
        ]
      },
      "sources": [
        "S4"
      ],
      "issues": [
        "SFT-I001"
      ],
      "experiments": [
        "SFT-E001",
        "SFT-E004"
      ],
      "decisions": [
        "SFT-D002"
      ],
      "tasks": [],
      "artifacts": [],
      "corrects": null
    },
    {
      "id": "SFT-H004",
      "date": "2026-09-06",
      "title": "补充锚点并修引用",
      "kind": "补录",
      "occurred_at": null,
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "fields": {
        "当天结论": [
          "补充锚点并修引用"
        ],
        "完成的工作": [
          "359 Pro / held-out 分析落地；引用审计修订。[[S2]][[S7]]"
        ],
        "新证据与结果": [
          "359 Pro / held-out 分析落地；引用审计修订。[[S2]][[S7]]"
        ],
        "判断与决策变化": [
          "补录仅归纳原记录；详细决定与历史结果见关联原文。"
        ],
        "问题变化": [
          "关联问题用于追踪背景；这里不推断当天已解决。"
        ],
        "论文变化": [
          "具体稿件版本见论文栏目；本条历史补录不推断当天改稿。"
        ],
        "下一步与交接": [
          "这是历史记录，当前行动以最新概览和接手包为准。"
        ]
      },
      "sources": [
        "S2",
        "S7"
      ],
      "issues": [
        "SFT-I002",
        "SFT-I003",
        "SFT-I103"
      ],
      "experiments": [],
      "decisions": [
        "SFT-D002",
        "SFT-D003"
      ],
      "tasks": [],
      "artifacts": [],
      "corrects": null
    },
    {
      "id": "SFT-H005",
      "date": "2026-09-06",
      "title": "保留测量限制",
      "kind": "补录",
      "occurred_at": null,
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "fields": {
        "当天结论": [
          "保留测量限制"
        ],
        "完成的工作": [
          "同条件重测停止，未完成结果不加入结论。[[S2]]"
        ],
        "新证据与结果": [
          "同条件重测停止，未完成结果不加入结论。[[S2]]"
        ],
        "判断与决策变化": [
          "补录仅归纳原记录；详细决定与历史结果见关联原文。"
        ],
        "问题变化": [
          "关联问题用于追踪背景；这里不推断当天已解决。"
        ],
        "论文变化": [
          "具体稿件版本见论文栏目；本条历史补录不推断当天改稿。"
        ],
        "下一步与交接": [
          "这是历史记录，当前行动以最新概览和接手包为准。"
        ]
      },
      "sources": [
        "S2"
      ],
      "issues": [
        "SFT-I002",
        "SFT-I003",
        "SFT-I103"
      ],
      "experiments": [],
      "decisions": [
        "SFT-D002",
        "SFT-D003"
      ],
      "tasks": [],
      "artifacts": [],
      "corrects": null
    },
    {
      "id": "SFT-SITE-V2",
      "date": "2026-10-04",
      "title": "研究工作台资料同步",
      "kind": "工作记录",
      "occurred_at": "2026-10-04T11:56:20+08:00",
      "observed_at": "2026-09-06",
      "recorded_at": "2026-10-04T11:56:20+08:00",
      "fields": {
        "当天结论": [
          "研究工作台上线资料整理；没有新增科研实验。"
        ],
        "完成的工作": [
          "建立日报、问题、实验、稿件与完整资料的关联入口。"
        ],
        "新证据与结果": [
          "公开现有来源，所有科研数字沿用各自观察时间。"
        ],
        "判断与决策变化": [
          "科研结论不因网站上线改变；网页用于恢复上下文。"
        ],
        "问题变化": [
          "当前问题按既有证据收录，没有自动关闭。"
        ],
        "论文变化": [
          "保留现有研究说明与PDF版本；未改写论文。"
        ],
        "下一步与交接": [
          "后续参与工作的AI用同步入口更新记录、验证并发布；仅阅读者无需写入。"
        ]
      },
      "sources": [],
      "issues": [],
      "experiments": [],
      "decisions": [],
      "tasks": [],
      "artifacts": [
        "CURRENT_PDF"
      ],
      "corrects": null
    },
    {
      "id": "SFT-BACKFILL-ANCHOR-CLOSURE",
      "date": "2026-10-04",
      "title": "补录已解决的教师锚点缺口",
      "kind": "补录",
      "occurred_at": null,
      "observed_at": "2026-09-06",
      "fields": {
        "当天结论": [
          "历史锚点在09-06已补充；本次只整理问题解决证据，没有新增实验。"
        ],
        "完成的工作": [
          "历史锚点在09-06已补充；本次只整理问题解决证据，没有新增实验。"
        ],
        "新证据与结果": [
          "历史锚点在09-06已补充；本次只整理问题解决证据，没有新增实验。"
        ],
        "判断与决策变化": [
          "历史锚点在09-06已补充；本次只整理问题解决证据，没有新增实验。"
        ],
        "问题变化": [
          "历史锚点在09-06已补充；本次只整理问题解决证据，没有新增实验。"
        ],
        "论文变化": [
          "历史锚点在09-06已补充；本次只整理问题解决证据，没有新增实验。"
        ],
        "下一步与交接": [
          "历史锚点在09-06已补充；本次只整理问题解决证据，没有新增实验。"
        ]
      },
      "sources": [
        "S2",
        "S4"
      ],
      "issues": [
        "SFT-I090"
      ],
      "experiments": [
        "SFT-E004"
      ],
      "decisions": [],
      "tasks": [],
      "artifacts": [],
      "corrects": null,
      "recorded_at": "2026-10-04T12:09:45+08:00"
    }
  ],
  "restricted": [
    {
      "id": "SEALED_TEST",
      "title": "封存测试与任务级标签",
      "path": "原项目的holdout/test材料",
      "availability": "未公开",
      "access": "在原环境按冻结协议与授权访问；发布流程不读取。"
    },
    {
      "id": "LARGE_ARTIFACTS",
      "title": "模型权重与完整轨迹",
      "path": "原项目模型/run目录",
      "availability": "体量过大，未发布",
      "access": "通过原项目文件清单、run身份及授权环境获取。"
    },
    {
      "id": "ACCOUNT_CREDENTIALS",
      "title": "账号凭据",
      "path": null,
      "availability": "不公开",
      "access": "使用已有连接或原环境授权，不写入接手文本。"
    }
  ],
  "manuscript": {
    "title": "What Agentic SFT Buys: Rapid Behavioral Convergence, Base-Conditional Capability Gains",
    "revision": "2026-09-06",
    "pdf_document": "CURRENT_PDF",
    "rows": [
      [
        "摘要 / 引言",
        "已有草稿",
        "主张与语料规模存在；最终范围仍需与 run 差异核对。[[S1]][[S4]]"
      ],
      [
        "相关工作",
        "已有草稿",
        "引用审计曾修订元数据和引用语境；不等于新颖性获得最终认证。[[S7]]"
      ],
      [
        "语料与实验设置",
        "已有草稿",
        "需把筛选变化、训练配置和异常处理讲成同一口径。[[S3]][[S5]]"
      ],
      [
        "行为、成本与结果",
        "已有草稿",
        "实测表与锚点存在；剂量、温度与 capability 措辞是重点复核项。[[S3]][[S4]]"
      ],
      [
        "讨论 / Limitations",
        "已核对",
        "现稿明确单教师、锚点范围、模型标注、仓库暴露、复合变体与单 rollout 限制。[[S5]]"
      ],
      [
        "附录 / 图表",
        "已有草稿",
        "A—E 在 main.tex 中接入；页面图直接采用主表数值。[[S1]][[S3]]"
      ],
      [
        "引用 / 审查 / PDF",
        "已核对",
        "当前 PDF 与引用审计报告存在；最终全篇论断与提交检查尚未由本页重新认证。[[S7]]"
      ]
    ],
    "notes": [
      "当前稿件已经包含摘要、八个主体章节、Limitations、伦理说明及 A—E 附录。不是最初只有框架的版本；本次核对确认章节和产物存在，但没有新执行全篇科研审查。[[S1]]",
      "计划里的投稿方向属于历史安排，网页不将其中的会议日期或截稿日当成已验证的当前规则。正式交付前应再次核对目标会议、正文页数、匿名和材料要求。完整 PDF 的总页数包含附录，不能当作主文页数。[[S2]]"
    ]
  },
  "background_file": "sft.json",
  "ingested_records": {
    "SFT-BACKFILL-ANCHOR-CLOSURE": "15b7fc2f2e39975177e515af2539da91d69d6667446116184ea8c071faecef8c"
  },
  "release_version": "v2-eb1a844c9b153986",
  "snapshot_recorded_at": "2026-10-04T12:09:45+08:00",
  "source_drift": []
}
